Artificial Intelligence-Augmented Knowledge Management in Public Administration: An Evidence-Based Taxonomy and Implementation Maturity Assessment

Public administrations are knowledge-intensive organizations. Their services depend on procedures, precedents, records, policy interpretations, and the accumulated expertise of public employees. Knowledge management (KM) provides mechanisms for creating, storing, sharing, and applying this institutional knowledge [1-3]. Yet public-sector KM is constrained by fragmented information systems, staff turnover, hierarchical structures, legal accountability, and uneven incentives for knowledge sharing [2, 4]. Artificial intelligence (AI) can address parts of this problem by classifying documents, extracting structured information, retrieving organizational knowledge, supporting decisions, and mediating access to complex repositories. The relevant question is therefore not whether AI is entering public administration, but which KM functions it actually augments, under what organizational conditions, and with what level of empirical maturity.

Existing reviews illuminate only parts of this intersection. Public-sector AI reviews describe applications and governance challenges but do not organize evidence around KM functions [5-7]. AI-KM reviews, by contrast, identify machine-learning, natural-language-processing, and generative-AI techniques across organizational settings, but their empirical base is mainly enterprise-oriented and does not distinguish public-sector prototypes from organizational deployments [8-12]. This separation matters because a model evaluated on government documents is not equivalent to a system owned, used, or evaluated by a public organization. It also matters because technical accuracy does not establish adoption, accountability, sustained use, or public value.

This review addresses that gap through a structured and auditable synthesis of direct public-sector AI-KM evidence. It asks two questions: RQ1, which AI capabilities are being applied to which KM functions in public-sector settings, and how mature is the supporting evidence? RQ2, which implementation and governance conditions distinguish experimental AI-KM systems from systems evaluated or used in operational public administration? The contribution is an evidence taxonomy with three analytical dimensions: KM function, AI capability, and implementation maturity. The review also develops a research agenda centered on organizational evaluation, governance-by-design, multilingual and resource-constrained deployment, and the transition from technical prototypes to sustained administrative use. Morocco and the wider Middle East and North Africa (MENA) are treated as an important research context rather than as evidence of mature deployment unless an identifiable public organization is shown to own, use, or evaluate the system.

3.1 Evidence maturity: A large corpus, but limited organizational deployment

The updated evidence base changes the interpretation of the field. Rather than a small literature dominated by reviews and policy sources, the final synthesis contains 170 source-validated direct evidence units. However, the expansion in corpus size does not imply comparable expansion in organizational maturity. Prototypes and experiments account for 116 units (68.2%), while 22 (12.9%) are conceptual or framework studies. Only 14 units (8.2%) are pilots, and 14 (8.2%) are operational deployments; four (2.4%) are policy/governance analyses that still meet the direct organizational-evidence rule. Thus, only 28 of 170 units (16.5%) provide pilot or operational evidence. The dominant empirical pattern is technical feasibility, not sustained administrative use. The empirical-maturity distribution is shown in Figure 2.

23.png

Figure 2. Empirical maturity of the 170 validated public-sector artificial intelligence-knowledge management (AI-KM) evidence units

The operational cases are nevertheless informative. The Library of the National Congress of Chile reports AI-supported linked legislative data, NLP tagging, and machine-learning services used in parliamentary knowledge work [14]. PandemIQ Llama is linked to a platform serving more than 100 governmental and multilateral public-health organizations, with reports being integrated into the World Health Organization's EIOS environment [15]. A regional-administration case in Italy reports productivity and work-redesign effects from generative-AI use, while retaining human validation requirements [16]. Conversely, the government-hospital ChatAI case shows that deployment is not equivalent to sustainability: sociotechnical misalignment, infrastructure change, regulation, and evolving work practices eventually contributed to discontinuation [17]. These cases make organizational fit a central analytical variable rather than an afterthought.

3.2 Artificial intelligence capabilities and knowledge management functions are concentrated rather than lifecycle-wide

The corpus is multi-label, so one evidence unit may involve several AI capabilities and KM functions. NLP/information extraction is the most common AI family (57 units; 33.5%), followed by LLM/generative AI (52; 30.6%), RAG (30; 17.6%), knowledge-graph/semantic AI (27; 15.9%), deep-learning/transformer models (25; 14.7%), and classical ML (18; 10.6%). Agentic or multi-agent approaches appear in only nine units (5.3%). This distribution does not support a claim that public administration has broadly progressed from classical ML to autonomous agents. Older and newer capability families coexist, and the more agentic end of the spectrum remains comparatively small.

The KM distribution is similarly uneven. Retrieval/access/question answering appears in 92 units (54.1%), application/decision support in 73 (42.9%), codification/structuring in 51 (30.0%), sharing/service delivery in 49 (28.8%), capture/extraction in 48 (28.2%), and governance/compliance/audit in 41 (24.1%). Storage/repository functions appear in 13 units (7.6%), while learning/evolution/tacit-knowledge functions appear in only nine (5.3%). The evidence therefore supports a narrower conclusion than the previous manuscript: AI is most visible in document- and retrieval-centered knowledge work, decision support, structuring, extraction, service delivery, and compliance. Evidence for the full KM lifecycle, especially tacit-knowledge development and organizational learning, remains limited. Table 3 summarizes these multi-label AI-capability and KM-function profiles.

Table 3. Artificial intelligence (AI)-capability and knowledge management (KM)-function profile of the final direct-evidence corpus

3.3 Recurring artificial intelligence-knowledge management interaction patterns

The co-occurrence matrix in Figure 3 reveals several recurring patterns. NLP is broadly distributed across capture, codification, retrieval, decision support, and governance. This reflects a mature class of public-sector tasks in which unstructured administrative text must be transformed into routable or analyzable information. Examples include a Serbian e-government question-answering system over public-administration documents [18] and a BERT-based Chinese government issue-handling system trained on 812,322 citizen inquiries to classify, characterize, and route cases to departments [19]. These systems are less visible than recent LLM assistants, but their organizational task is concrete, and their dependence on clearly bounded classification or retrieval functions can simplify validation.

tu_pian_3.png

Figure 3. Co-occurrence of major artificial intelligence (AI) capabilities and knowledge management (KM) functions in the validated direct-evidence corpus

Knowledge graphs and semantic AI are more closely associated with codification and retrieval, while LLMs and RAG are strongly associated with retrieval, decision support, and governance-oriented knowledge access. Public-service crowdsourcing work shows how AI-assisted processing can support collective KM and systems change [20]. In Austria, explainable AI was used to transform legal text into computational logic in a Ministry of Finance case, with domain experts involved in validation [21]. CaltransGPT illustrates the RAG pattern in transportation-agency institutional repositories [22]. DataGovAI applies multi-agent RAG to data-retention compliance for Utah state-government employees, with a supervisor agent validating cited responses [23]. A U.S. Air Force policy and logistics prototype similarly uses RAG over regulatory publications and includes evaluation by an Air Force officer [24]. These examples show why the unit of analysis should be the administrative knowledge task and organizational setting, not the novelty of the model alone.

3.4 Capability gains must be evaluated against governance and implementation risks

The field is often narrated as a technological progression from ML to LLMs, RAG, and agents. The evidence supports a different interpretation: each capability family changes the balance between flexibility and control. Classical classifiers can be narrow but comparatively easy to benchmark. Knowledge graphs improve explicit structure and traceability but require ontology and data-maintenance effort. LLMs broaden natural-language access but introduce hallucination, prompt sensitivity, privacy, and evaluation problems. RAG can ground generation in organizational sources [25], but retrieval error, stale or unauthorized documents, embedding leakage, and privacy attacks remain material risks [26]. Agentic systems add orchestration and task execution, which can improve workflow coverage but also increase the number of decisions, tools, and state transitions that must be governed.

The implementation indicators reinforce this point. Real administrative data were present in 110 units (64.7%), but public-sector users were explicitly involved in only 39 (22.9%); another 66 units described intended or mixed user involvement. Human oversight was explicit or clearly retained in 130 units (76.5%). These figures suggest that the literature is comparatively good at demonstrating AI on realistic public-sector information but much weaker at evaluating how public employees use, contest, correct, or institutionalize the systems. The evidence therefore does not justify a universal prescription for sovereign, on-premise, or agentic deployment. Data location, compute architecture, model openness, and automation level should instead be treated as context-dependent design choices shaped by sensitivity, law, service criticality, connectivity, and organizational capability. Table 4 summarizes these implementation indicators, while Table 5 compares the main capability-risk trade-offs.

Table 4. Implementation indicators in the final direct-evidence corpus

Table 5. Capability-risk comparison for public-sector artificial intelligence-knowledge management (AI-KM)

3.5 Technical feasibility is ahead of organizational and public-value evidence

A central result of the review is the separation between technical performance and organizational evidence. Many prototypes report accuracy, retrieval, relevance, or generation metrics, yet fewer examine changes in work quality, service time, accountability, adoption, sustained use, or public value. The Thai People Map and Analytics Platform conversational agent translates natural-language questions into database queries for government analytics, but remains a bounded prototype [27]. CaltransGPT reports grounding and citation-oriented evaluation on a small query set [22]. DataGovAI compares RAG configurations for compliance support [23]. These are meaningful technical results, but none should be described as evidence that the corresponding organization has achieved mature, routine, organization-wide AI-KM.

The most valuable operational studies therefore do more than report model performance. They show integration with work, users, institutional repositories, or governance arrangements [14-17]. The hospital ChatAI case is particularly important because discontinuation itself is evidence: successful technical functionality did not prevent later sociotechnical misalignment [17]. Future evaluations should consequently report at least four layers: technical quality; human-task performance; organizational adoption and workflow effects; and governance/public-value outcomes. Without this separation, review conclusions risk rewarding technically sophisticated prototypes while overlooking simpler systems that have stronger institutional fit.

3.6 Morocco and Middle East and North Africa: Research context, not demonstrated maturity

The Morocco/MENA search streams were included to prevent a geographically generic evidence map from being mistaken for a universal one. Previous work by the authors identified adoption challenges, guidelines, and success factors for KM systems in Moroccan public administration [28, 29]. More recent Moroccan research demonstrates relevant technical capability, including NLP and LLM classification of Darija requests for administrative documents [30]. However, the final source-level validation demoted that study from direct public-sector implementation evidence because no Moroccan administration was shown to own, deploy, or organizationally evaluate the proposed chatbot. This distinction is deliberate. Strategy, technical readiness, language resources, and public-service motivation are important context, but they are not substitutes for implementation evidence.

The revised conclusion for Morocco is therefore more cautious and more useful. The available evidence supports a research opportunity at the intersection of multilingual public-service knowledge, KM capability, governance, and constrained infrastructure; it does not support a claim that Moroccan administrations have already converged on a mature AI-KM deployment model. Priority studies should identify specific administrative workflows, involve public employees, evaluate Arabic/Darija/French knowledge access under realistic data-governance constraints, and measure sustained organizational outcomes.

3.7 Positioning against prior reviews

The evidence taxonomy differs from earlier reviews in three ways. First, public-sector KM reviews established organizational and knowledge-governance issues before the current generative-AI wave [4]. Second, public-sector AI reviews mapped application areas and adoption constraints without using KM functions as the principal analytical structure [5-7, 31]. Third, AI-KM reviews examined technique-process relationships across organizations but did not apply a strict organizational public-sector evidence rule or a maturity distinction between prototype, pilot, and operation [8-12, 32-34]. The present review therefore adds more than a geographic filter: it links AI capability to KM function while independently coding whether the evidence represents technical experimentation or actual administrative use. That additional dimension changes the substantive conclusion from "AI is diffusing across KM" to "AI is concentrated in a subset of public-sector knowledge functions, and deployment maturity is much lower than technical activity suggests."

3.8 Research agenda and limitations

Five priorities follow from the evidence. First, the field needs longitudinal implementation studies that track use, correction, abandonment, and institutionalization rather than only benchmark performance. Second, evaluations should connect model metrics to administrative outcomes such as case-processing time, consistency, auditability, knowledge reuse, employee workload, and service quality. Third, governance should be evaluated as an implemented control system: access-aware retrieval, provenance, logging, human decision rights, update governance, and responsibility should be tested rather than merely recommended. Fourth, multilingual and resource-constrained administrations require dedicated evidence on local-language retrieval, offline or hybrid operation, model maintenance, and data-residency constraints. Fifth, comparative studies should examine when narrower NLP, rule-based, or semantic systems outperform generative or agentic alternatives on cost, reliability, explainability, and organizational fit.

The review has limitations. Screening and coding were conducted by a single reviewer; the 814-record consistency check measures intra-reviewer stability, not inter-rater reliability. Source verification was not uniformly based on full text because access varied across publications, so the final claims are limited to what could be verified at source level. English and French dominated the formal screening, which may underrepresent local-language public-sector work. Finally, the field is changing rapidly, and the August 2026 search is a time-bounded snapshot. The analytical categories should therefore be treated as an evidence map that can be updated rather than as a fixed maturity model.

During the preparation of this manuscript, the authors used a generative AI assistant to support drafting, language refinement, and editorial revision. All AI-assisted text was reviewed, edited, and validated by the authors, who take full responsibility for the accuracy, interpretation, references, and final content of the article.